i4h-workflow-dataset-mimic

Use trajectory mimic augmentation, which clones existing episodes from an HDF5 recording while applying small action and state noise. This expands the dataset for policy training without new teleoperation, though it perturbs actions and states rather than visuals.

Point the IN variable at an existing HDF5 recording, then run the mimic run.sh script with the environment id, episode count, noise standard deviation, and output path. Verify the output episode count equals the source demos plus the requested new episodes.

Yes. After mimic completes, chain the dataset conversion skill to convert the augmented HDF5 to LeRobot format with h264 video, then load the LeRobot visualization skill pointed at the converted dataset directory to inspect episodes.

This error means the source recording contains zero successful episodes, typically because teleop was launched but no demos were saved. Record successful source demonstrations first; do not substitute an older run unless the user explicitly requests it.

Mimic perturbs only actions and states, not visual observations, and it augments an existing recording rather than capturing new episodes. The output state and action dimensions must match the source recording, so the environment id must match the original data.